[https://nvbugs/6402048][test] Skip MXFP8 CUTLASS MoE with non-128-aligned per-shard intermediate

[xxi-nv]

Description

MXFP8CutlassFusedMoEMethod.create_weights hard-asserts intermediate_size_per_partition % 128 == 0 for its int32 UE8M0 scale-factor packing. When moe_tp_size > 1 shards a non-128-aligned intermediate axis — e.g. Qwen1.5-MoE-A2.7B intermediate_size=1408 with TTP/DTP moe_tp_size=4 gives per-shard 352 — the assertion raises and test_configurable_moe_multi_gpu -k "CUTLASS and FP8 and not FP8_BLOCK_SCALES and not W4A8" fails (nvbugs/6402048). The post-merge selector was matching MXFP8 via the FP8 substring even though it was intended for plain FP8 QDQ.

Changes

1. should_skip_cutlass — add QuantAlgo.MXFP8 to tp_alignment_quants, so non-128-aligned per-shard MXFP8 configs are skipped at parameter-generation time, matching W8A16 / NVFP4 / W4A8_AWQ.
2. waives.txt — remove the now-redundant SKIP entry (the case no longer reaches collection).
3. l0_dgx_b200.yml post-merge — add and not MXFP8 to the CUTLASS and FP8 ... selector so it stays on plain FP8 QDQ, as its comment intends.
4. l0_dgx_b200.yml pre-merge — add an explicit CUTLASS and MXFP8 and not W4A8 entry so aligned MXFP8 configs get real coverage.

Scope check

• Qwen1.5-MoE (i=1408) + TTP/DTP — per-shard 352, now skipped (the failing case). EP modes (DEP/TEP, unsharded 1408) still run.
• DeepSeek-V4 (i=2048), boundary (i=512) — all parallel modes still run (per-shard 512 / 128 aligned).
• GPT-OSS (h/i=2880) — already skipped by the generic 128-alignment check in get_quick_skip_reason.
• plain FP8 — FP8QDQFusedMoEMethod has no alignment assertion; unaffected.
• H100 / B300 — H100 (SM90) rejects MXFP8 via can_implement (MXFP8 needs SM100/103); B300 uses detailed test-ids that don't include MXFP8. Neither needs changes.

Test Coverage

unittest/_torch/modules/moe/test_moe_module.py::test_configurable_moe_multi_gpu

• Post-merge: plain FP8 QDQ (MXFP8 excluded).
• Pre-merge: aligned MXFP8 CUTLASS configs now explicitly covered.

Notes

Supersedes repair-bot PR #15846 (in-kernel padding fix); this takes the lighter test-skip route consistent with the sibling non-128-aligned quant algos.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR modifies the should_skip_cutlass TP-per-shard alignment logic to include QuantAlgo.MXFP8 in the alignment quant set, updates related comments, and removes a corresponding SKIP waiver entry for a multi-GPU MoE test that filtered on CUTLASS and FP8 conditions.

Changes

MoE CUTLASS Skip Logic and Waiver Update

Layer / File(s)	Summary
TP alignment skip condition and waiver removal tests/unittest/_torch/modules/moe/moe_test_utils.py, tests/integration/test_lists/waives.txt	QuantAlgo.MXFP8 added to tp_alignment_quants in should_skip_cutlass with revised comments documenting the 128-aligned per-shard requirement and MXFP4 auto-padding for W4A8_MXFP4_MXFP8; corresponding SKIP waiver entry for test_configurable_moe_multi_gpu removed.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#15825: Directly counteracts this PR by having previously added the same waiver entry that is now removed.
• NVIDIA/TensorRT-LLM#15850: Both PRs modify tests/integration/test_lists/waives.txt for skipped MoE test configurations.

Suggested reviewers: taylor-yb-lee, nvchenghaoz

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is concise and accurately describes the main change to MXFP8 CUTLASS MoE test skipping.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The PR description clearly explains the issue, fix, scope, and test coverage, and matches the required template well.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[xxi-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #57119 [ run ] triggered by Bot. Commit: 8f8088e Link to invocation

[tensorrt-cicd]

PR_Github #57119 [ run ] completed with state SUCCESS. Commit: 8f8088e
/LLM/main/L0_MergeRequest_PR pipeline #45906 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[xxi-nv]

/bot run --disable-fail-fast

[xxi-nv]

/bot run --disable-fail-fast

[xxi-nv]

/bot run --disable-fail-fast

[xxi-nv]

/bot run --disable-fail-fast

[xxi-nv]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #57548 [ run ] triggered by Bot. Commit: 2171b35 Link to invocation

[tensorrt-cicd]

PR_Github #57548 [ run ] completed with state SUCCESS. Commit: 2171b35
/LLM/main/L0_MergeRequest_PR pipeline #46277 completed with status: 'SUCCESS'

CI Report

Link to invocation